Dear Mr. Carlo Catoni,

A little bit python code with openCV library could help you write your own program as per your requirement. Can you attach a sample video so that I can share some more information (or may be the specific code to run) as per your need?

Best wishes..

With regards...

Bhaskar

Dear Carlo Catoni,

This is a paper of my colleagues. In the paper you can find information on previous works, and they will probably help you on available tools if you contact them.

https://www.researchgate.net/publication/310426189_The_Evaluation_of_Detectors_and_Descriptors_on_Determination_of_Semen_Cell

Conference Paper The Evaluation of Detectors and Descriptors on Determination...

Hello. Your question is about tracking and thereby counting sperm cells in a video sequence. There are a number of open source libraries, like the one Bhaskar has mentioned above. Be careful though, there are many practical problems that prevent simple algorithms such as nearest neighbour, on which OpenCV and other open source code is typically based, from giving accurate results.

The hidden assumptions in this type of algorithm are that there is good SNR (a high detection probability), a low false alarm rate (few spurious measurements), good separation between the objects being tracked (no occlusion) and that they move in straight lines, at least when viewed at the frame rate of the video system. Another assumption relates to the nature of the object being tracked, which is an extended object in your case, but for which the centroid will usually be the part that is actually tracked.

The machine learning community has only recently (since the 1990s) discovered tracking as a discipline and it proposing various methods for solving this rather old problem. The first tracking algorithms arose in the 1960s in connection with military radar systems and since that time, scores of different techniques have been proposed to handle tracking scenarios of various types and complexities. The simplest were based on approximate Kalman filters called alpha-beta filters, using a nearest-neighbour association rule. The most complicated are based on multi-hypothesis filtering algorithms that include provision for manoeuvres and so on. I have not read that many papers from the machine learning community that demonstrate a good understanding of the tracking literature, although there are some notable exceptions. A good one is Cox and Hingorani's paper from IEEE PAMI in 1996.

You would do well to read the paper below, which was published before particle filters and set-based estimators (like PHD) became popular (at least in the research sense). It is useful to remember that tracking is one of those areas where it is quite hard to say what approach works best because of the difficulty of defining what "best" means. There are a large number of possible failure modes of a tracking algorithm, not just mean square error (track initiation time, track loss, track swaps, etc).

Please upvote if you like this answer, or even better, reference my paper ;-)

Article Taxonomy of multiple target tracking methods

Hello and thanks everybody.

Given my almost null capability of programming, I was hoping to receive some information about a tracking software that already can do this out of shelf. I found ImageJ that can do something, but not all I need..

Thanks anyway